Biology subjects

Wei, K. Y.

Publications and source records attributed to Wei, K. Y..

4 recordsLinked to original sources

bioRxiv2026

TaxoFormer: Hierarchical Transformer for Predicting the Full Taxonomic Lineage of Protein Sequences

Predicting labels in massive, hierarchically structured output spaces is a core challenge in machine learning. In this work, we use the problem of predicting the full taxonomic lineage of a protein from its sequence as a case study for this challenge. We introduce TaxoFormer, an architecture whose primary contribution is a structured tokenization scheme that losslessly represents the entire NCBI phylogenetic tree, a graph with over 1.3 million nodes using a compact vocabulary of just 15,000 tokens. By coupling a pre-trained ESM-2 model with an autoregressive decoder and training with a standard cross-entropy objective, we test the hypothesis that a simple generative objective is sufficient to learn complex, latent structure when the output space is explicitly modeled. We show that this approach is highly effective: on a dataset of 188 million proteins, the model not only achieves accurate lineage prediction but also implicitly learns a continuous, phylogenetically-structured latent space. This work provides a scalable, alignment-free method for taxonomic annotation and demonstrates that explicitly modeling the structure of a complex output space is a powerful mechanism for learning meaningful representations.2

synthetic biology↗

bioRxiv2025

A GENERALIZED PROTEIN DESIGN ML MODEL ENABLESGENERATION OF FUNCTIONAL DE NOVO PROTEINS

Traditional protein design is fundamentally constrained by known sequences and folds. To break free from these limitations, we introduce a new alternative: designing proteins directly from plain-language specifications. To achieve this, we trained MP4, a transformer-based model that maps natural language prompts to protein sequences, on a dataset of 3.2 billion points and 138k tokens. In a benchmark of 96 prompts representing a wide array of functions and contexts, MP4 excelled by simultaneously improving on three key metrics: sequence realism, predicted fold quality, and alignment to the requested function. This high performance is particularly significant as it was achieved using only text as input which is a major departure from other models. Experimental validation confirmed our computational predictions: two de novo designs were experimentally shown to be both expressible and thermostable, with high-resolution crystallography (1.30 [A] and 1.77 [A]) ultimately revealing one to possess a paradigm-shifting novel fold. Functionally, the designs were also active, demonstrating both ATP binding and hydrolysis in vitro. This work demonstrates the realization of natural-language intent as functional proteins that express, crystallize, and catalyze. Although the underlying approach is still in early development with incomplete coverage and controllability, MP4 delivers a profound impact: it lowers the barrier to protein design and vastly expands the space for creative exploration in molecular programming.

biochemistry↗

bioRxiv2022

Towards generalizable prediction of antibody thermostability using machine learning on sequence and structure features

Over the last three decades, the appeal for monoclonal antibodies (mAbs) as therapeutics has been steadily increasing as evident with FDAs recent landmark approval of the 100th mAb. Unlike mAbs that bind to single targets, multispecific biologics (bsAbs) with their single-chain variable fragment (scFv) modules have garnered particular interest owing to the advantage of engaging distinct targets. Despite their exquisite specificity and affinity, the relatively poor thermostability of these scFv modules often hampers their development as a potential therapeutic drug. In recent years, engineering antibody sequences to enhance their stability by mutations has gained considerable momentum. As experimental methods for antibody engineering are time-intensive, laborious, and expensive, computational methods serve as a fast and inexpensive alternative to conventional routes. In this work, we show two machine learning methods - one with pre-trained language models (PTLM) capturing functional effects of sequence variation, and second, a supervised convolutional neural network (CNN) trained with Rosetta energetic features - to better classify thermostable scFv variants from sequence. Both these models are trained over temperature-specific data (TS50 measurements) derived from multiple libraries of scFv sequences. In this work, we show that a sufficiently simple CNN model trained with energetic features generalizes better than a pre-trained language model on out-of-distribution (blind) sequences (average Spearman correlation coefficient of 0.4 as opposed to 0.15). Further, we demonstrate that for an independent mAb with available thermal melting temperatures for 20 experimentally characterized thermostable mutations, these models trained on TS50 data could identify 18 residue positions and 5 identical amino-acid mutations showing remarkable generalizability. Our results suggest that such models can be broadly applicable for improving the biological characteristics of antibodies. Further, transferring such models for alternative physico-chemical properties of scFvs can have potential applications in optimizing large-scale production and delivery of mAbs or bsAbs.

bioinformatics↗

bioRxiv2019

De novo design of proteins with two isoenergetic but structurally divergent ground states

The plasticity of naturally occurring protein structures, which can change shape considerably in response to changes in environmental conditions, is critical to biological function. While computational methods have been used to de novo design proteins that fold to a single state with a deep free energy minima (Huang et al., 2016), and to reengineer natural proteins to alter their dynamics (Davey et al., 2017) or fold (Alexander et al., 2009), the de novo design of closely related sequences which adopt well-defined, but structurally divergent structures remains an outstanding challenge. Here, we design closely related sequences (over 94% identity) that can adopt two very different homotrimeric helical bundle conformations -- one short ([~]66 [A] height) and the other long ([~]100 [A] height) -- reminiscent of the conformational transition of viral fusion proteins (Ivanovic et al., 2013; Podbilewicz, 2014; Skehel and Wiley, 2000). Crystallographic and NMR spectroscopic characterization show that both the short and long state sequences fold as designed. We sought to design bistable sequences for which both states are accessible, and obtained a single designed protein sequence that populates either the short state or the long state depending on the measurement conditions. The design of sequences which are poised to adopt two very different conformations sets the stage for creating large scale conformational switches between structurally divergent forms.

biochemistry↗